Chapter 2: Iterative Combinatorial Auctions
|
|
- Kevin Harris
- 5 years ago
- Views:
Transcription
1 Chapter 2: Iterative Combinatorial Auctions David C. Parkes 1 Introduction Combinatorial auctions allow bidders to express complex valuations on bundles of items, and have been proposed in settings as diverse as the allocation of floor space in a new condominium building to individual units (Wired 2000) and the allocation of take-off and landing slots at airports (Smith, Forward). Many applications are described in Part V of this book. The promise of combinatorial auctions (CAs) is that they can allow bidders to better express their private information about preferences for different outcomes and thus enhance competition and market efficiency. Much effort has been spent on developing algorithms for the hard problem of winner determination once bids have been received (Sandholm, Chapter 14). Yet, preference elicitation has emerged as perhaps the key bottleneck in the real-world deployment of combinatorial auctions. Advanced clearing algorithms are worthless if one cannot simplify the bidding problem facing bidders. Preference elicitation is a problem both because of the communication cost of sending bids to the auction and also because of the cost on bidders to determine their valuations for different bundles. The problem of communication complexity can be addressed through the design of careful bidding languages, that provide expressive but concise bids (Nisan Chapter 9). Non-computational approaches can also be useful, such as defining the good and bundle space in the right way in the first place (Pekeč and Rothkopf Chapter 16). However, even well-designed sealed-bid auctions cannot address the problem of hard valuation problems because they preclude the use of feedback and price 1
2 discovery to focus bidder attention. There are an exponential number of bundles to value in CAs. Moreover, the problem of valuing even a single bundle can be difficult in many applications of CA technology. For instance, in the airport landing slot scenario (see Ball, Donohue and Hoffman Chapter 20) we should imagine that airlines are solving local scheduling, marketing, and revenue-management problems to determine their values for different combinations of slots. Iterative combinatorial auctions are designed to address the problem of costly preference elicitation that arises due to hard valuation problems. An iterative CA allows bidders to submit multiple bids during an auction and provides information feedback to support adaptive and focused elicitation. For example, an ascending price auction maintains ask prices and allows bidders to revise bids as prices are discovered. Significantly, it is often possible to determine an efficient allocation without bidders reporting, or even determining, exact values for all bundles. In contrast, any efficient sealed-bid auction requires bidders to report and determine their value for all feasible bundles of goods. This ability to mitigate the preference elicitation problem is a central concern in iterative CA design. But there are also a number of less tangible yet still important benefits: Iterative CAs can help to distribute the computation in an auction across bidders through the interactive involvement of bidders in guiding the dynamics of the auction. Some formal models show the equivalence between iterative CAs and decentralized optimization algorithms (Parkes and Ungar 2000a, de Vries, Schummer, and Vohra 2003). Iterative CAs can address concerns about privacy because bidders only need to reveal partial and indirect information about their valuations. 1 Transparency is another practical concern in CAs. In the high-stakes world of wireless spectrum auctions, the Federal Communications Commission 2
3 (FCC) has been especially keen to ensure that bidders can verify and validate the outcome of an auction. Although mathematically elegant, the VCG outcome can be difficult to explain to bidders, and validation requires the disclosure and verification of many bids, both losing and winning. Thus, even as readily describable implementations of sealed-bid auctions, iterative CAs can offer some appeal (Ausubel and Milgrom 2002). The dynamic exchange of value information between bidders, that is enabled within iterative CAs, is known to enhance revenue and efficiency in single item auctions with correlated values (Milgrom and Weber 1982). Although little is known about the design of iterative CAs for correlated value problems, one should expect iterative CAs to retain this benefit over their sealed-bid counterparts. Certainly, correlated value settings exist: consider the wireless spectrum auctions in which valuations are in part driven by underlying population demographics and shared technological realities. Yet, even with all these potential advantages iterative CAs offer new opportunities to bidders for manipulation. The biggest challenge in iterative CA design is to support incremental and focused bidding without allowing new strategic behavior to compromise the economic goals of efficiency or optimality. For instance, one useful design paradigm seeks to implement auctions in which straightforward bidding (truthful demand revelation in response to prices) is an ex post equilibrium. This equilibrium is invariant to the private information of bidders, so that straightforward bidding is a best response whatever the valuations of other bidders. Steps can also be taken to minimize opportunities for signaling through careful control of the information that can be shared between bidders during an auction. Finally, the benefits of iterative auctions disappear when bidders choose to strategically delay bidding activity until the last rounds of an auction. Activity 3
4 rules (Milgrom 2000) can be used to address this stalling and promote meaningful bidding during the early rounds of an auction. The existing literature on iterative CAs largely focuses on the design of efficient auctions. Indeed, there are no known optimal (i.e. revenue-maximizing) general-purpose CAs, iterative or otherwise. As such, the canonical VCG mechanism (see Chapter 1) has guided the design of many iterative auctions. 2 We focus mainly on price-based approaches, in which the auctioneer provides ask prices to coordinate the bidding process. We also consider alternative paradigms, including decentralized protocols, proxied auctions in which a bidding agent elicits preference information and automatically bids using a predetermined procedure, and direct-elicitation approaches. In outline, Section 2 defines competitive equilibrium (CE) prices for CAs, which may be non-linear and non-anonymous in general. Connections between CE prices, the core of the coalitional game, and the VCG outcome are explained. Section 3 describes the design space of iterative CAs. Section 4 discusses price-based auctions, providing a survey of existing price-based CAs in the literature and a detailed case study of an efficient ascending price auction. Section 5 considers some alternatives to price-based design. Section 6 closes with a brief discussion of some of the open problems in the design of iterative combinatorial auctions, and draws some connections with the rest of this book. 2 Preliminaries Let G = {1,..., m} denote the set of items, and assume a private values model with v i (S) 0 to denote the value of bidder i I = {1,...,n} for bundle S G. Note that set I does not include the seller. We assume free-disposal, with v i (T) v i (S) for all T S, and normalization, with v i ( ) = 0. Let V denote the set of bidder valuations. Bidders are assumed to have quasi-linear utility (we also use payoff interchangeably with utility), with utility u i (S,p) = v i (S) p for bundle S at price p 0. This assumes the absence of any budget constraints. 4
5 Further assume that the seller has no intrinsic value for the items. The efficient combinatorial allocation problem (CAP) solves: max v i (S i ) S=(S 1,...,S n) i I s.t. S i S j =, i,j [CAP(I)] Let S denote the efficient allocation. Also, we write CAP(I \ i) to denote the combinatorial allocation problem without bidder i. 2.1 Competitive Equilibrium Prices We can consider a hierarchical structure for ask prices in CAs: Linear. Prices p j 0, for j G, define additive prices on bundles, with p(s) = j S p j. Non-linear. Prices, p(s) 0, for S G, allow p(s) p(s 1 ) + p(s 2 ), for some S = S 1 S 2 and S 1 S 2 =. Non-linear and Non-anonymous. Prices p i (S) 0, allow discriminatory pricing, with p i (S) p i (S) for bidder i i, in addition to non-linear prices. In the following definitions we adopt p i (S) for notational convenience. We intend to allow (but not require) with this notation non-linear and non-anonymous prices. For instance, linear prices p j can be considered to induce prices p i (S) = j S p j for bundle S and bidder i. Competitive equilibrium prices extend the concept of Walrasian equilibrium prices to a CA. Let π i (S,p) = v i (S) p i (S) denote bidder i s payoff from bundle S at prices p and Π s (S,p) = i I p i(s i ) denote the seller s revenue from allocation S at prices p. 5
6 Definition 1 (Competitive Equilibrium). Prices, p, and allocation S = (S 1,...,S n) are in competitive equilibrium (CE) if: π i (Si,p) = max [v i(s) p i (S),0] i (1) S G Π s (S,p) = max p i (S i ) (2) S Γ i I where Γ denotes the set of all feasible allocations. A competitive equilibrium (p,s ) is such that allocation S maximizes the payoff of every bidder and the seller given prices. Allocation S is said to be supported by prices p in CE. Theorem 1. Allocation S is supported in competitive equilibrium if and only if S is an efficient allocation. This welfare theorem follows from a simple linear-programming (LP) duality argument for suitably extended LP formulations of the CAP (Bikhchandani and Ostroy 2002, also Chapter 8). Moreover, CE prices always exist for the CAP. For instance, prices p i = v i trivially satisfy the CE conditions. The main new element in CAs is that these CE prices must sometimes be non-linear and non-anonymous. Bikhchandani and Ostroy also show an equivalence between the core of the coalitional game and the set of CE prices. All core outcomes can be priced, and all CE prices correspond to core payoffs. Many iterative CAs are designed to converge to CE prices, and as such it is important to characterize classes of valuations for which linear, and non-linear but anonymous, CE prices exist. We will also see that it is necessary that an efficient CA must determine enough information about bidder valuations to define a set of CE prices, and necessary that a Vickrey auction determines enough information to define a set of universal CE prices. For the existence of linear CE prices, it is sufficient (and almost necessary) 3 that valuations satisfy a goods are substitutes property (Kelso and Crawford 1982, Gul 6
7 and Stacchetti 1999). This substitutes condition is defined indirectly, with respect to a demand set: D i (p) = {S : π i (S,p) max T G π i(t,p),π i (S,p) 0,S G}, (3) which includes all bundles that maximize a bidder s payoff at the prices. Definition 2 (Goods are Substitutes). Valuation v i satisfies goods are substitutes if for all linear prices p,p such that p p (component-wise), and all S D i (p), there exists T D i (p ) such that {j S : p j = p j } T. The goods are substitutes (or simply substitutes) condition requires that a bidder will continue to demand items that do not change in price as the price on other items increases. Substitutes valuations include unit-demand valuations with v i (S) = max j S {v ij } for all S and value v ij on item j in isolation, but preclude the possibility of items with complementary values (Lehmann, Lehmann, and Nisan 2001). Conditions for the existence of non-linear but anonymous CE prices are less well-understood, but sufficient conditions presented in Parkes (2001) (Theorem 4.7) include supermodular valuations, single-minded bidders that value a particular bundle, and bidders with safe valuations such that each pair of bundles with positive value to a bidder share at least one item. Consider, for example, a bidder in the FCC spectrum auction that definitely needs lower Manhattan, along with as many of the geographically neighboring licenses as possible. 2.2 Minimal Competitive Equilibrium Prices In fact, many iterative CAs are designed to converge to minimal CE prices. This can be useful for two reasons. First, minimal CE prices on bundles in the efficient allocation correspond to VCG payments for a restricted class of valuations. In this case, we say that the CE prices support the VCG payments. Termination with CE prices that support VCG payments brings straightforward bidding into an ex post equilibrium. Second, Ausubel and Milgrom (2000, also Chapter 3) show that 7
8 implementing minimal CE prices (corresponding to buyer-optimal core outcomes) avoids the problems that can occur with the VCG auction when VCG payments are not supported with minimal CE prices. Definition 3 (Minimal CE Prices). Minimal CE prices minimize the seller s total revenue Π s (S,p) on the efficient allocation S across all CE prices. A bidder s payment in the VCG mechanism is always less than or equal to the payment by that bidder at any CE price (Bikhchandani and Ostroy 2002). Thus, minimal CE prices always provide an upper-bound on VCG payments. Moreover, a bidder s VCG payment is equal to the CE price on her efficient bundle in some CE (Parkes and Ungar 2000b). A characterization in terms of the coalitional value function explains when the VCG can be supported simultaneously to all bidders in the minimal CE. Let w(l) for L I denote the coalitional value for a subset L of bidders, equal to the value of the efficient allocation for CAP(L). The buyers are substitutes (BAS) condition requires, w(i) w(i \ K) [w(i) w(i \ i)], K I (BAS) i K Theorem 2. (Bikhchandani and Ostroy 2002) A buyers are substitutes (BAS) coalitional value function is necessary and sufficient to support the VCG payments in competitive equilibrium. In particular, the VCG payments are implemented in the minimal CE (or buyer-optimal core) when BAS holds, and buyer-optimal core payoffs are unique exactly when BAS holds. A number of ascending price CAs can only terminate with minimal CE prices given a slightly stronger condition, that of a buyer-submodular (BSM) coalitional value function: w(l) w(l \ K) [w(l) w(l \ i)], K L, L I (BSM) i K 8
9 Bikhchandani and Ostroy (Chapter 8) refer to BSM as buyers are strong substitutes. Clearly, a BSM coalitional value function also satisfies BAS. But there are cases for which values satisfy BAS but not BSM (see Ausubel and Milgrom 2002, Section 7, for example). Interestingly, substitutes valuations implies BSM and is almost necessary. Roughly, if at least one bidder does not satisfy substitutes then one can construct substitutes valuations for other bidders such that the coalitional value function fails BSM. See Ausubel and Milgrom (Chapter 1) for further discussion. Thus, the same conditions for the existence of a linear price equilibrium are sufficient and almost necessary for the existence of some price equilibrium (although perhaps non-linear and non-anonymous) that supports the Vickrey outcome Universal Competitive Equilibrium Prices Experiments have suggested that BAS can often fail in realistic settings for CAs. 5 In these cases the VCG payments are not supported in any price equilibrium. We can still design price-based CAs by characterizing a stronger condition on CE prices that implies enough information to determine VCG payments from these prices. For this, we restrict attention to the universal CE prices (Parkes and Ungar 2002, Mishra and Parkes 2004). Definition 4 (Universal CE Prices). Prices p are universal Competitive Equilibrium (UCE) prices if: a) Prices p are CE prices. b) Prices p i are CE prices for CAP(I \ i), meaning they support some efficient allocation in CAP(I \ i), for all bidders i. where p i = (p 1,...,p i 1,p i+1,...,p n ). In words, prices are UCE when an efficient allocation for the restricted allocation problem without bidder i is supported with prices p i, for each bidder i removed 9
10 in turn. Thus, UCE prices are CE prices in the main economy and in every marginal economy. Note that UCE prices need not require that the same allocation is supported in every marginal economy. The prices must support some efficient allocation in each marginal economy. 6 UCE prices always exist, for example p i = v i, for all bidders i, are UCE prices. Moreover, a universal price equilibrium provides sufficient information about bidder valuations to compute the VCG outcome. Theorem 3. (Parkes and Ungar 2002) Given a UCE with prices p uce and an efficient allocation S, the VCG payment to bidder i is computed as: p vcg,i = p uce,i (S i ) [Π I(p uce ) Π I\i (p uce)] (4) where Π L (p) = max S Γ i L p i(s i ) for bidders L I. In the special case when prices are equal to valuations then this adjustment is equivalent to the standard definition of VCG payments. 2.4 Informational Requirements Both CE and UCE prices have a central role in the preference elicitation problem. First, any auction that implements an efficient allocation must determine a set of CE prices. Second, any auction that implements the Vickrey outcome must determine a set of UCE prices. Segal (Chapter 11) provides an extended discussion. Since the VCG auction is basically unique amongst the class of efficient auctions that take a zero payment from losing bidders (Ausubel and Milgrom, Chapter 1), these equivalences confirm the central role of prices in developing iterative CAs. Theorem 4. (Parkes 2002, Nisan and Segal 2003) A combinatorial auction realizes the efficient allocation if and only if the auction also realizes a set of CE prices and an allocation supported in the price equilibrium. 10
11 This result requires a technical condition of privacy-preservation, which precludes bidders from making their valuations contingent on the valuations of other bidders (e.g. my value for A is at least bidder 2 s value for A ). 7 Theorem 5. (Parkes and Ungar 2002, Lahaie and Parkes 2004b) A combinatorial auction realizes the VCG outcome if and only if the auction also realizes a set of UCE prices and an allocation supported in the price equilibrium of the main economy. That UCE prices provide sufficient information was first proved in Parkes and Ungar (2002). The necessary direction is due to Lahaie and Parkes (2004b). It is important to realize that the CE and UCE prices referenced in these results may only be realized implicitly and are not necessarily explicitly constructed in the auctions. Considering minimal CE prices in particular, Mishra and Parkes (2004) note that minimal CE prices are universal iff BAS holds. In general, UCE prices are greater than the minimal CE prices because they must consider competition in the marginal economies in addition to the main economy. The informational equivalence between the efficient outcome and the problem of discovering CE prices leads to a (largely negative) characterization of the worst-case communication complexity and preference-elicitation requirements of any efficient CA, iterative or otherwise (Segal, Chapter 11). On the other hand, iterative CAs are designed to have good elicitation properties on typical instances, while sealed-bid auctions must suffer the worst case every time. Moreover, this price equivalence suggests the central role of prices in the design of iterative CAs. Any protocol to determine the VCG outcome must (implicitly) determine UCE prices, so why not construct protocols to converge directly to UCE prices? We return to this theme in Section 4. 11
12 2.5 Examples The following examples illustrate the concept of CE and UCE prices and also serve to illustrate the principle that it is often unnecessary to receive complete information about bidder valuations to determine the Vickrey outcome. For each example, we define a space of valuations (that contain the true valuations) that provides sufficient information to determine the Vickrey outcome. The information is minimal we call this a minimal information set in the sense that no relaxed constraints on valuations are sufficient to pin down the Vickrey outcome. Example 2.1 Consider a single-item auction with three bidders and values (10, 8, 6). The efficient allocation assigns the item to bidder 1, and the Vickrey payment is $8. Prices 10 p 8 are all in CE, and p = $8 is the unique anonymous UCE price. Notice that the UCE price must be at least $8 to satisfy CE condition (1) for bidder 2 in CAP({1,2,3}) but no greater than $8 to satisfy the same condition for bidder 2 in CAP({2,3}). The CE prices define a minimal information set, ˆV 1, defined as the subset of valuations that satisfy constraints {v 1 p,v 2 p,v 3 p,10 p 8}. UCE prices imply additional information {v 2 = 8,v 3 8}, which together with ˆV 1 is a minimal information set for the VCG outcome. Notice that an ascending price (i.e. English) auction can elicit this information if bidders 1 and 2 bid up the price to just above 8, at which point the auction terminates. Bidder 3 can remain silent. Example 2.2 Consider a combinatorial allocation problem with items {A,B} and 5 bidders (see Figure 2.1). The efficient allocation allocates A to bidder 1 and B to bidder 2 for a total value of 70. The VCG payments are p vcg,1 = 30 (70 65) = 25 and p vcg,2 = 40 (70 55) = 25. Figure 2.1 (b) illustrates an information set on 12
13 bidder valuations, that is sufficient to compute the VCG outcome and minimal in the sense that no constraints can be relaxed. The following prices are UCE for any valuation in this set: p(a) = 25,p(B) = 25,p(AB) = 25 to bidders {1,2,4,5} and prices p 3 (A) = 20,p 3 (B) = 20,p 3 (AB) = 40 to bidder 3. In fact, these prices are also minimal CE prices and the discount computed in Eq. 4 is zero for bidders 1 and 2, and BAS is satisfied (because of the presence of bidders 4 and 5). Without these bidders, the BAS condition fails and the VCG payments become p vcg,1 = 0 and p vcg,2 = 20, which can be computed from UCE prices p 1 = (20,0,20),p 2 = (0,40,40) and p 3 = (0,20,40). Additional information is needed from bidder 2 in this variation. 3 The Design Space for Iterative Combinatorial Auctions The design space for iterative CAs is larger than for one-shot auctions. Important considerations include the design of information feedback to bidders and rules to guide the submission of bids. Cramton (Chapter 4) provides an in-depth discussion of many of these issues in the design of simultaneous ascending price auctions. Let the state of an auction include all the information that is sufficient to define the future dynamics of the auction. For example, the state of an auction can define the ask prices, the provisional allocation, and also the bid improvement rules as they apply to particular bidders. Briefly, we can consider the role of the following design features: Timing issues. Iterative auctions may be continuous, allowing bids to be submitted at any time with continual updates to the current provisional allocation and prices. Alternatively, iterative auctions may be discrete, or round-based, with the state updated periodically and with bidders provided with an opportunity to revise bids between rounds. Continuous auctions can promote faster propagation of feedback information to bidders and help to quickly focus elicitation. However, 13
14 continuous combinatorial auctions can be infeasible because the winner-determination problem must be resolved whenever a new bid is submitted. Continuous auctions also lead to high monitoring and participation costs for bidders. In comparison, discrete auctions allow an auctioneer to publish a schedule for rounds in the auction and bidders can plan when to allocate time to refine their values and bids. Information feedback. Information feedback about the state of an auction can include information about the bids submitted and also aggregate information, such as price feedback and the current provisional allocation, to guide bidding. Information hiding is also possible, for example with rounding to limit the potential for signaling between bidders and with limited and discriminatory reporting of bid information. Information feedback policies make a tradeoff between serving the goal of providing effective bid guidance and minimizing the opportunity for collusion and other forms of manipulation through signaling and coordination. Bidding Rules. Ask prices are a common form of bid improvement rule, placing a lower-bound on the allowable bid price on a bundle. Bid improvement rules can also require a minimal percentage improvement over the current highest bid on a bundle, or over the total revenue in the next round given current bids. Activity rules (Milgrom 2000) introduce further restrictions, such as requiring that a bidder bids for a decreasing market share as prices increase during an auction. Ausubel, Cramton and Milgrom (Chapter 5) provide an extended discussion of bid-improvement and activity rules. Activity rules were introduced in the early FCC wireless spectrum auctions and proved important. 8 Decisions about appropriate rules are often guided by a tradeoff between providing expressiveness so that bidders can follow straightforward bidding strategies, while promoting early information 14
15 exchange between bidders and limiting the opportunity for bidders to wait and snipe at the end of an auction. Computational considerations also matter, for example linear prices can simplify the problem facing bidders in an auction (Kwasnica, Ledyard, Porter, and DeMartini 2004) but can be expensive to compute (Hoffman 2001). Termination Conditions. Auctions may close at a fixed deadline, perhaps with an opportunity for a final sealed-bid round of bidding (sometimes called a proxy round). Alternatively, auctions can have a rolling closure with the auction kept open while one or more losing bidders continue to submit competitive bids. Fixed deadlines are useful in settings in which bidders are impatient and unwilling to wait a long time for an auction to terminate. However, fixed deadlines tend to require stronger activity rules to prevent the auction reducing to a sealed-bid auction with bids delayed until the final round. In comparison, rolling closure rules have been shown to promote early and sincere bidding. 9 Bidding Languages. A bid can be a complex object and expressed in terms of logical connectives (Nisan, Chapter 9). One popular bidding language is exclusive-or (XOR), in which bid (p 1,S 1 ) xor (p 2,S 2 ) xor... xor (p l,s l ) has semantics I will buy at most one of these bundles at the stated bid price. Another popular language is additive-or (OR) bidding languages, in which bid (p 1,S 1 ) or (p 2,S 2 ) or... or (p l,s l ) has semantics I will buy one or more of these bundles at the stated bid price. Bidding languages can also place constraints on the bid prices, for example by requiring click-box bidding in which bidders must submit bids from a menu. 10 The expressiveness of a bidding language in an iterative CA must be considered together with the opportunity to refine bids during an auction. For instance, a language that is additive-or on items is not expressive in a 15
16 one-shot CA but becomes expressive in an ascending auction when bidders can decommit from bids. 11 Bidding languages are often designed to support straightforward bidding with bidders able to state the bundle that maximizes their surplus in response to prices in each round. Proxy agents. Proxy agents provide a still richer interface for iterative CAs (Parkes and Ungar 2000b, Ausubel and Milgrom 2002). Bidders can provide direct value information to an automated bidding agent that bids on their behalf within an auction. The bidder-to-proxy language should allow a bidder to express partial and incomplete information, to be refined during the auction, in order to realize the elicitation and price discovery benefits of an iterative auction. Proxy agents can query a bidder actively when they have insufficient information to submit bids. Proxy agents can also facilitate faster convergence with rapid automated proxy rounds interleaved with bidder rounds. Mandatory proxy agents can be useful in restricting the strategy space available to bidders. One concern in the design of proxy auctions is to determine when to allow proxy information to be revised and to determine the degree of consistency to enforce across revisions. An additional concern is that of trust and transparency since the bidding activity is transferred to automated agents. 4 Price-Based Iterative Combinatorial Auctions Many iterative CAs are price based and provide ask prices to guide bidding. In this section we survey some of these auction designs. We limit our attention to auctions designed for valuations that are rich enough to include the substitutes valuations. As such, we exclude the assignment model in which bidders have unit-demand for items. See Bikhchandani and Ostroy (Chapter 8) for a taxonomy that includes this case. 16
17 All the auctions that we discuss share the same high level structure: In each round the auctioneer announces ask prices and a provisional allocation and requests new bids from bidders. The bids are used to formulate a new winner-determination problem and update the provisional allocation, and also to adjust ask prices and test for termination. Table 2.1 provides a summary of the characteristics of some well-known auctions, stating properties for straightforward (non-strategic) bidding. For the cases in which an auction terminates with the VCG outcome this assumption is justified in an ex post equilibrium but otherwise one should expect incentives for demand reduction. The auctions are described in terms of the structure of the price space, the bidding language, and the method used to update prices. 17
18 Name Valuations Price structure Bidding Price Update Outcome language method KC substitutes non-anon items OR-items greedy CE SAA substitutes items OR-items greedy CE GS substitutes items XOR minimal min CE 12 Aus substitutes items single greedy a VCG ibundle; Ascending-proxy b BSM non-anon bundles XOR greedy VCG... general min CE dvsv BSM non-anon bundles XOR minimal VCG Clock-proxy BSM items (+ proxy) c XOR greedy VCG... general min CE RAD general items OR LP-based AkBA general anon bundles XOR LP-based ibea general non-anon bundles XOR greedy d VCG MP general non-anon bundles XOR minimal d VCG Table 2.1: Price-Based Combinatorial Auctions. Formal properties are stated for straightforward bidding, and with the most general class of valuations for which the property holds. Notation in the Outcome column indicates that no formal properties have been established. Notes: a Aus traces n + 1 trajectories. b Ascending-proxy dynamics are identical to ibundle(3), although ascending-proxy emphasizes a sealed-bid proxy auction form. c Clock-proxy is a hybrid design, with a linear-price clock auction followed by a sealed-bid ascending-proxy auction. d Ascending price while the auction is open, followed by a downwards adjustment after termination. Abbreviations: KC (Kelso and Crawford 1982) SAA (Milgrom 2000) GS (Gul and Stacchetti 2000) Aus (Ausubel 2002) ibundle (Parkes and Ungar 2000a) Ascending-proxy (Ausubel and Milgrom 2002) dvsv (de Vries, Schummer, and Vohra 2003) Clock-proxy (Ausubel and Milgrom, Chapter 5) RAD (Kwasnica et al. 2003) AkBA (Wurman and Wellman 2000) ibea (Parkes and Ungar 2002) MP (Mishra and Parkes 2004) We see a wide variety of prices, from simple prices on items (linear prices) to non-anonymous prices on bundles (non-anonymous and non-linear). In addition, the auctions vary in the bids that a bidder can submit in each round: OR-items, an additive-or bid for multiple items; XOR, an exclusive-or bid for multiple bundles; single, a bid on a single bundle in each round; OR, an additive-or bid for multiple 18
19 bundles. The XOR language has emerged as the definitive choice in recent designs. The price-update methods, which characterize the rules by which prices are computed in each round, are broken down as follows: Greedy update: The price is increased on some arbitrary set (perhaps all) of the over-demanded items or bundles. Minimal update: The price is increased on a minimal set of overdemanded items, or based on the bids from a set of minimally undersupplied bidders. LP-based: A linear program, formulated to find prices that are good approximations to CE prices given current bids, is used to adjust prices. For linear prices, Demange, Gale and Sotomayor (1986) in the assignment model and later Gul and Stacchetti (2000) for substitutes define a minimal update in terms of increasing the prices on a minimal overdemanded set of items. 13 Minimal price updates are adopted to drive prices towards minimal CE prices. de Vries, Schummer and Vohra (2003) generalize this to define updates in terms of minimally undersupplied bidders 14 and define a minimal update for general CAs. All bidders in a minimally undersupplied set face higher prices on the bundles for which they submitted a bid. RAD and AkBA adopt LP-based price updates and adjust prices to find good approximations to CE prices given current bids and the current provisional allocation. RAD seeks linear and anonymous prices while AkBA seeks non-linear but anonymous price approximations. Formal convergence properties have not been proved for RAD or AkBA, although RAD reduces to a simultaneous ascending price auction for substitutes valuations. The auctions that are able to implement the VCG outcome (for instance, Aus for substitutes and dvsv for BSM coalitional values) are interesting because they bring straightforward bidding into an equilibrium. Straightforward bidding is a 19
20 best response, whatever the valuations of other bidders, as long as the other bidders also follow a straightforward (perhaps untruthful) bidding strategy. This ex post equilibrium concept is useful because it places no requirements on the knowledge that bidders have of the valuations of other bidders. Winning bidders pay their final bid price in all auctions except Aus, ibea and MP. Aus allows for (n + 1) restarts and uses information elicited along each trajectory to determine the final payments. ibea and MP terminate with UCE prices, at which point final payments are determined through downwards adjustments. Auction clock-proxy (Ausubel, Cramton and Milgrom Chapter 5) is a hybrid auction. The first stage maintains item prices and runs an ascending-clock CA (see also Porter, Rassenti, and Smith (2003)). This stage is used for price discovery and can be considered to construct approximate linear CE prices. The second stage is sealed-bid, with bids from the first stage combined with additional bids that must be consistent with bids from the clock phase. 4.1 Insufficiency of Simple Prices It is interesting to consider what form of prices are necessary to implement efficient ascending CAs. Gul and Stacchetti (2000) first addressed this question, in the setting of substitutes valuations. The authors provide a formal definition of an ascending CA, but limit attention to linear and anonymous prices. They show that there exists no ascending VCG auction with linear and anonymous prices for substitutes valuations. The auction due to Ausubel (2002) lies outside of this negative characterization because it uses n + 1 price trajectories. Recently, Mishra and Parkes (2004) used the UCE-based price characterization to demonstrate that efficient ascending CAs require both non-anonymous and non-linear prices, even for this case of substitutes valuations. The authors exhibit instances for which only non-anonymous and non-linear UCE prices exist. As for sufficiency, auctions dvsv and ibundle are examples of ascending VCG auctions for substitutes valuations that maintain these rich prices. 20
21 However, de Vries, Schummer, and Vohra (2003) extend the definition of ascending CAs in Gul and Stacchetti (2000) to allow for non-anonymous and non-linear prices and obtain a negative result. When at least one bidder has a non-substitutes valuation an ascending CA cannot implement the VCG outcome even when the other bidders are restricted to substitutes and even with non-anonymous and non-linear prices. Auctions ibea and MP lie outside of this negative characterization because they allow a final downwards adjustment to determine final prices. Thus, with substitutes values but simple prices we must accept auctions with multiple trajectories or non-monotonic adjustments. Moreover, although rich prices extend the reach of ascending CAs to substitutes values we still need to accept multiple trajectories or non-monotonic adjustments to handle richer preferences than substitutes. 4.2 Primal-Dual Auction Design Many traditional combinatorial optimization problems can be solved with primal-dual algorithms. A primal-dual approach uses linear-programming (LP) duality to formulate an optimization problem as a satisfaction problem. Strong LP duality states that a pair of feasible primal and dual solutions are optimal if and only if they satisfy complementary slackness (CS) conditions. We provide a brief review of LP theory at the end of this chapter, and refer the reader to Papadimitriou and Steiglitz (1998) for a textbook treatment. In fact, primal-dual theory also provides a useful conceptual framework for the design of iterative price-based CAs. Prices are interpreted as a feasible dual solution and the provisional allocation is interpreted as a feasible primal solution. Bids provide sufficient information to formulate and solve restricted primal and dual problems, the winner-determination and price-update problems respectively (see Figure 2.2). For further discussion of this idea, see Parkes (2001), de Vries, Schummer and Vohra (2003) and Bikhchandani and Ostroy (Chapter 8). 21
22 Straightforward bidding is first assumed, and later justified by termination with VCG payments. The winner-determination problem uses information implicit in bids to compute a feasible solution that minimizes the violation of the CS conditions, and price updates adjust the dual solution towards an optimal dual solution. 15 CS conditions have an exact equivalence with conditions (1) and (2) required for CE prices, and are satisfied on termination of an auction. Constructively, primal-dual auction design requires the following steps: 1. Formulate an LP for the CAP that is integral, such that its optimal solution is a feasible allocation. The dual problem should allow convergence to UCE prices, or to minimal CE prices that support VCG payments in the case of BAS valuations. 2. Provide bidders with a bidding language that is expressive for straightforward bidding, and formulate a winner-determination problem to compute a feasible primal solution that minimizes the violation of CS conditions as represented in bids. 3. Terminate when the provisional allocation and ask prices satisfy CS conditions (and thus represent a CE), and also satisfy any additional conditions that are necessary to compute the VCG payments at termination (e.g. UCE conditions or minimal CE prices). Otherwise, adjust prices to make progress towards an optimal dual solution that satisfies these conditions. The characterization of VCG payments in terms of minimal CE and UCE prices suggests two methods to adjust towards the VCG outcome. The methods are illustrated in Figure 2.3, which considers the price on bundles, S 1 and S 2, allocated to bidders 1 and 2 in the efficient outcome. In case (a), the coalitional value function satisfies BSM and the VCG payments are supported at the minimal CE prices. Ascending CAs (such as dvsv) can 22
23 converge monotonically to these prices and the VCG outcome. In case (b), the coalitional value function satisfies neither BSM not BAS. Although each bidder s VCG payment is supported in some minimal CE there is no single CE that supports the VCG payment to both bidders simultaneously. As illustrated, ascending CAs such as ibea and MP can still converge monotonically to UCE prices from which the VCG outcome can be determined in a final adjustment. The next section presents a case study of primal-dual methods to the design and analysis of the ibundle auction. 16 In Section 4.4 we return to the auctions in Table 2.1, and discuss each in a little more detail. 4.3 Case Study: ibundle We will focus on variation ibundle(2), in which prices are non-linear but anonymous. This variation is efficient with straightforward bidding and an additional requirement that bidder strategies satisfy a safety property. Later, we also briefly describe ibundle(3), which employs non-linear and non-anonymous prices and is efficient without the safety condition. The interested reader is referred to Parkes and Ungar (2000a) and Parkes (2001) for additional details, including a description of ibundle(d), which blends ibundle(2) and ibundle(3) and allows for dynamic price discrimination decisions to be made during the auction. In what follows, we will use ibundle to refer to variation ibundle(2) unless otherwise stated. ibundle(2): Anonymous Prices ibundle maintains ask prices on bundles and a provisional allocation and proceeds in rounds, indexed t 1. In each round a bidder can submit XOR bids on bundles. In general the bid price on a bundle must be at least the ask price. Bidders must resubmit bids on any bundle that they are winning in the current provisional allocation but can bid at the same price on such a bundle even if the ask price has since increased. A bidder can also bid at ǫ less than the ask price 23
24 when making a last-and-final bid, at which point she can no longer improve her price. Equivalently, one can simply retain all bids from previous rounds. A bid at, or above, the current ask price is said to be competitive, and a bidder is competitive if she submits at least one competitive bid. The winner-determination problem in each round is to compute a provisional allocation to maximize the seller s revenue given bids, with at most one bundle selected from the XOR bid of each bidder. Let B i denote the bids from bidder i, and p bid,i (S) denote the bid price on bundle S B i. Winner determination can be formulated as the following mathematical program: s.t. max x i (S)p bid,i (S) x i (S) i I S B i x i (S) 1, i (5) S B i x i (S) 1, j (6) i I S B i :j S x i (S) {0,1}, i, S B i Constraint (5) restricts the seller to selecting at most one bid from each bidder. Constraint (6) ensures the allocation is feasible. Ties are broken first to favor the allocation from the previous round and then to maximize the number of winning bidders. ibundle terminates when each competitive bidder receives a bundle in the provisional allocation. Otherwise, prices are increased, by ǫ above the bid price on all bundles that receive a bid from some losing bidder in the current round and the new allocation and prices are provided as feedback to bidders. Prices on other bundles are implicitly adjusted to satisfy free disposal, although only bundles that receive losing bids need to be explicitly quoted. On termination the provisional allocation becomes the final allocation, and bidders pay their final bid prices. ibundle maintains feasible primal and dual solutions to an extended LP formulation of CAP and terminates with a CE outcome that satisfies CS 24
25 conditions. The proof technique is inspired by Bertsekas (1987) analysis of the AUCTION algorithm for the special case of unit-demand valuations. Given ask prices, p i (S), to bidder i we define ǫ straightforward bidding in terms of an ǫ-demand set, ǫ-ds, which is: ǫd i (p i ) = {S : v i (S) p i (S) + ǫ max S (v i (S ) p i (S ),0), S G} (ǫ-ds) In words, bidders state in their bid all bundles that come within ǫ of maximizing their surplus given prices in each round. This reduces to straightforward bidding for a small enough ǫ. Definition 5 (Safety). The competitive bundles in the ǫ-demand set of each losing bidder in each round are non-disjoint, i.e. each pair of bundles shares at least one item. For example, losing bids {(ABC, $100),(CDE, $50)} from a single bidder satisfy safety, while losing bids {(ABC, $100),(DE, $50)} from a single bidder fail the safety condition. Theorem 6. (Parkes and Ungar 2000a) ibundle(2) terminates with an allocation that is within 3min(n,m)ǫ of the efficient solution for ǫ-straightforward bidding strategies and with bid safety. The first step of the proof is to introduce an extended LP formulation (LP 2 ) for CAP due to Bikhchandani and Ostroy (2002, see also Chapter 8). LP 2 is integral when the safety condition holds for straightforward bidding. The dual formulation (DLP 2 ) has variables that correspond to anonymous and non-linear prices. Let K denote the set of feasible partitions. For example, (A, B, C) and (AB, C) are feasible partitions for items ABC. Variable y(k) = 1 will indicate that the allocation must be restricted to bundles in partition k K. For example, if partition (AB, C) is selected then the only valid allocations are those in which 25
26 AB goes to some bidder and C to another bidder. We have: s.t. max x i (S),y(k) S G i I i I x i (S)v i (S) [LP 2 ] x i (S) 1, i S G x i (S) y(k), S y(k) 1 k K k K:S k x i (S),y(k) 0, i,s,k min π i + Π s [DLP 2 ] π i,p(s),π s i I s.t. π i + p(s) v i (S), i,s Π s p(s) 0, k S k π i,p(s),π s 0, i,s Dual variable p(s) can be interpreted as the ask price on bundle S. Then, optimal πi = max S{v i (S) p(s),0} defines the maximal payoff to bidder i across all bundles given prices, and optimal Π s = max k K S k p(s) defines the maximal revenue to the seller across all partitions given prices. This is also the maximal revenue across all allocations because prices are anonymous. The dual problem sets prices to minimize the sum of the maximal payoff to each bidder and the maximal revenue to the seller. Optimal dual prices will correspond to CE prices whenever the primal LP is integral. Interpret the provisional allocation and ask prices in a round of ibundle(2) as defining a feasible primal and a feasible dual solution (denoted ˆx,ŷ, ˆπ i, ˆp, and ˆΠ s ). We can now establish termination with CS conditions for straightforward bidding strategies. The first primal CS condition is: ˆx i (S) >0 ˆπ i + ˆp(S) = v i (S), i,s (CS-1) 26
27 This states that any bundle allocated to bidder i must maximize her payoff across all bundles at the prices. Condition (CS-1) is approximately satisfied in every round because the provisional allocation is selected with respect to bids, which are in turn drawn from ǫ demand sets. Formally, a relaxed form of condition (CS-1) holds, with ˆx i (S) > 0 ˆπ i + ˆp(S) v i (S) + 2ǫ, for all i and S. The second primal CS condition is: ŷ(k) > 0 ˆΠ s ˆp(S) = 0, k (CS-2) S k This states that the provisional allocation must maximize the seller s payoff (i.e. revenue) given the prices, across all feasible allocations and irrespective of bids received from bidders. Bundle S has a strictly positive price if it is greater than the price on every bundle contained in S. Then, (CS-2) follows from properties (P1) and (P2), which are maintained in each round of the auction: (P1) All bundles with strict positive prices receive a bid from some bidder in every round. (P2) One or more of the revenue-maximizing allocations in every round can be constructed from bids from different bidders. Formally, (P1) follows because one can show that a losing bidder will continue to bid for S in the next round, even at the higher price. Property (P2) follows from the safety property, which prevents a single bidder from causing the price to increase on a pair of disjoint bundles. This is why we need the safety condition. Combining (P1) and (P2), and together with ǫ-ds, we get a relaxed formulation of (CS-2), with ŷ(k) > 0 ˆΠ s S k ˆp(S) min(m,n)ǫ, for all partitions k K. Dual CS condition (CS-3), states: ˆπ i > 0 ˆx i (S) = 1, i (CS-3) S G 27
28 In words, every bidder with positive payoff for some bundle at the current prices must receive a bundle in the provisional allocation. (CS-3) is satisfied for all bidders that receive bundles in a particular round, but not for the losing bidders that are still competitive. However, (CS-3) holds for every bidder on termination because at this point ǫ DS = for all losing bidders. (CS-3) and (CS-1) are equivalent to CE condition (1) and (CS-2) together with an additional requirement that a provisional allocation is always selected is equivalent to CE condition (2). Finally, we obtain an upper-bound on the worst-case efficiency error of ibundle, in terms of the minimal bid increment ǫ. First, sum the approximate (CS-1) condition over all bidders in the final allocation, and substitute ˆπ i = 0 for bidders not in the allocation by (CS-3). This gives: ˆπ i v i (Ŝi) ˆp(Ŝi) + 2min(m,n)ǫ (7) i I i I i I ˆΠ s + ˆπ i v i (Ŝi) + 3min(m,n)ǫ (8) i I i I where Eq. (7) follows because an allocation can include no more bundles than there are items or bidders, and Eq. (8) is by substitution of the ǫ-approximate (CS-2) condition. The LHS of Eq. (8) is the value of the final dual solution, and the first-term on the RHS is the value of the final primal solution. Now, ˆΠ s + i ˆπ i w(i), (the value of the optimal primal) by LP weak duality, and therefore w(i) ˆΠ s + i ˆπ i i v i(ŝi) + 3min(m,n)ǫ. A complete proof must also show termination. The basic idea is to assume the auction never terminates and prove that a bidder must eventually submit a bid at a price above her valuation, assuming finite values and a finite number of items, from which we get a contradiction with straightforward bidding. 28
29 ibundle(3): Non-anonymous Prices ibundle(3) is the variation of ibundle in which each bidder faces non-anonymous prices in every round. The dynamics of ibundle(3) with straightforward bidding are identical to that of Ausubel and Milgrom s (2002) ascending-proxy auction, although ascending-proxy is not described in price terms. ibundle(3) is efficient for straightforward bidding with general values. Moreover, the auction will terminate with VCG outcomes for BSM coalitional value functions. Let p t ask,i (S) denote the ask prices to bidder i in round t. Initially, p1 ask,i (S) = 0 for all bundles S and all bidders. Bids are received, and the winner determination problem solved, as in ibundle(2). Then, for each bidder not in the provisional allocation, the price to that bidder is increased by the minimal bid increment, ǫ > 0, above her bid price on all bundles submitted in that round, and adjusted for free-disposal. It is now quite immediate to establish that ibundle(3) terminates in CE with straightforward bidding. The prices faced by a bidder in round t are parameterized by πi t 0, which can be interpreted as the maximal payoff to the bidder in that round. The ask price on bundle S in round t is defined as: p t ask,i (S) = max(0,v i(s) π t i) (9) Initially, πi 1 = max S{v i (S)}, for all i, and the price is zero on all bundles. The payoff πi t decreases monotonically during the auction and prices monotonically increase. The ǫ-ds for bidder i in round t includes every bundle for which v i (S) πi t, and increases monotonically across rounds. Eventually, when πt i is less than ǫ the prices on each bundle are within ǫ of her value and she will bid for every bundle with positive value in her ǫ-ds. 17 Condition (CS-1) holds trivially in each round and (CS-3) holds at termination, just as in ibundle(2). In addition, (CS-2) holds in each round because of the special structure of prices: every bundle with a strict positive price receives a bid in a bidder s ǫ-ds. This does not require the safety condition. 29
An Introduction to Iterative Combinatorial Auctions
An Introduction to Iterative Combinatorial Auctions Baharak Rastegari Department of Computer Science University of British Columbia Vancouver, B.C, Canada V6T 1Z4 baharak@cs.ubc.ca Abstract Combinatorial
More informationMulti-Item Vickrey-Dutch Auction for Unit-Demand Preferences
Multi-Item Vickrey-Dutch Auction for Unit-Demand Preferences The Harvard community has made this article openly available. Please share how this access benefits you. Your story matters. Citation Accessed
More informationOn the Robustness of Primal-Dual Iterative Combinatorial Auctions
Vol. 00, No. 0, Xxxxx 0000, pp. 000 000 issn 0000-0000 eissn 0000-0000 00 0000 0001 INFORMS doi 10.1287/xxxx.0000.0000 c 0000 INFORMS On the Robustness of Primal-Dual Iterative Combinatorial Auctions Stefan
More informationExpressing Preferences with Price-Vector Agents in Combinatorial Auctions: A Brief Summary
Expressing Preferences with Price-Vector Agents in Combinatorial Auctions: A Brief Summary Robert W. Day PhD: Applied Mathematics, University of Maryland, College Park 1 Problem Discussion and Motivation
More informationSpectrum Auction Design
Spectrum Auction Design Peter Cramton Professor of Economics, University of Maryland www.cramton.umd.edu/papers/spectrum 1 Two parts One-sided auctions Two-sided auctions (incentive auctions) Application:
More informationIterative Combinatorial Auctions: Theory and Practice
From: AAAI-00 Proceedings. Copyright 2000, AAAI (www.aaai.org). All rights reserved. Iterative Combinatorial Auctions: Theory and Practice David C. Parkes and Lyle H. Ungar Computer and Information Science
More informationModels for Iterative Multiattribute Procurement Auctions
Models for Iterative Multiattribute Procurement Auctions The Harvard community has made this article openly available. Please share how this access benefits you. Your story matters. Citation Published
More informationAn Experimental Comparison of Linear and Non-Linear Price Combinatorial Auctions
An Experimental Comparison of Linear and Non-Linear Price Combinatorial Auctions Tobias Scheffel, Alexander Pikovsky, Martin Bichler, Kemal Guler* Internet-based Information Systems, Dept. of Informatics,
More informationA Modular Framework for Iterative Combinatorial Auctions
A Modular Framework for Iterative Combinatorial Auctions SÉBASTIEN LAHAIE Yahoo Research and DAVID C. PARKES Harvard University We describe a modular elicitation framework for iterative combinatorial auctions.
More informationActivity Rules and Equilibria in the Combinatorial Clock Auction
Activity Rules and Equilibria in the Combinatorial Clock Auction 1. Introduction For the past 20 years, auctions have become a widely used tool in allocating broadband spectrum. These auctions help efficiently
More informationA Modular Framework for Iterative Combinatorial Auctions
A Modular Framework for Iterative Combinatorial Auctions The Harvard community has made this article openly available. Please share how this access benefits you. Your story matters Citation Lahaie, Sébastien,
More informationSponsored Search Markets
COMP323 Introduction to Computational Game Theory Sponsored Search Markets Paul G. Spirakis Department of Computer Science University of Liverpool Paul G. Spirakis (U. Liverpool) Sponsored Search Markets
More informationA Kernel-Based Iterative Combinatorial Auction
Proceedings of the Twenty-Fifth AAAI Conference on Artificial Intelligence A Kernel-Based Iterative Combinatorial Auction Sébastien Lahaie Yahoo! Research New York, NY 10018 lahaies@yahoo-inc.com Abstract
More informationAn Ascending Price Auction for Producer-Consumer Economy
An Ascending Price Auction for Producer-Consumer Economy Debasis Mishra (dmishra@cae.wisc.edu) University of Wisconsin-Madison Rahul Garg (grahul@in.ibm.com) IBM India Research Lab, New Delhi, India, 110016
More informationSoftware Frameworks for Advanced Procurement Auction Markets
Software Frameworks for Advanced Procurement Auction Markets Martin Bichler and Jayant R. Kalagnanam Department of Informatics, Technische Universität München, Munich, Germany IBM T. J. Watson Research
More informationCOMP/MATH 553 Algorithmic Game Theory Lecture 8: Combinatorial Auctions & Spectrum Auctions. Sep 29, Yang Cai
COMP/MATH 553 Algorithmic Game Theory Lecture 8: Combinatorial Auctions & Spectrum Auctions Sep 29, 2014 Yang Cai An overview of today s class Vickrey-Clarke-Groves Mechanism Combinatorial Auctions Case
More informationLecture 7 - Auctions and Mechanism Design
CS 599: Algorithmic Game Theory Oct 6., 2010 Lecture 7 - Auctions and Mechanism Design Prof. Shang-hua Teng Scribes: Tomer Levinboim and Scott Alfeld An Illustrative Example We begin with a specific example
More informationNew Developments in Auction Theory & Practice. Jonathan Levin Cowles Lunch, March 2011
New Developments in Auction Theory & Practice Jonathan Levin Cowles Lunch, March 2011 Introduction Developments in auction theory and practice Rapid expansion in theory and practice of multi item auctions,
More informationCoherent Pricing of Efficient Allocations in Combinatorial Economies
Coherent Pricing of Efficient Allocations in Combinatorial Economies Wolfram Conen XONAR GmbH Wodanstr. 7 42555 Velbert, Germany E-mail: conen@gmx.de Tuomas Sandholm Carnegie Mellon University Computer
More informationGame Theory: Spring 2017
Game Theory: Spring 2017 Ulle Endriss Institute for Logic, Language and Computation University of Amsterdam Ulle Endriss 1 Plan for Today This and the next lecture are going to be about mechanism design,
More informationRevaluation of Bundles by Bidders in Combinatorial Auctions
Revaluation of Bundles by Bidders in Combinatorial Auctions Soumyakanti Chakraborty Indian Institute of Management Calcutta, Joka, Diamond Harbour Road, Kolkata 700104 fp07004@iimcal.ac.in Anup Kumar Sen
More informationCS364B: Frontiers in Mechanism Design Lecture #1: Ascending and Ex Post Incentive Compatible Mechanisms
CS364B: Frontiers in Mechanism Design Lecture #1: Ascending and Ex Post Incentive Compatible Mechanisms Tim Roughgarden January 8, 2014 1 Introduction These twenty lectures cover advanced topics in mechanism
More informationNon-decreasing Payment Rules in Combinatorial Auctions
Research Collection Master Thesis Non-decreasing Payment Rules in Combinatorial Auctions Author(s): Wang, Ye Publication Date: 2018 Permanent Link: https://doi.org/10.3929/ethz-b-000260945 Rights / License:
More informationNote on webpage about sequential ascending auctions
Econ 805 Advanced Micro Theory I Dan Quint Fall 2007 Lecture 20 Nov 13 2007 Second problem set due next Tuesday SCHEDULING STUDENT PRESENTATIONS Note on webpage about sequential ascending auctions Everything
More informationTraditional auctions such as the English SOFTWARE FRAMEWORKS FOR ADVANCED PROCUREMENT
SOFTWARE FRAMEWORKS FOR ADVANCED PROCUREMENT A range of versatile auction formats are coming that allow more flexibility in specifying demand and supply. Traditional auctions such as the English and first-price
More informationmany autonomous individuals in human societies, it is perhaps natural to look to economic principles to design coordination mechanisms in computationa
Chapter 1 Introduction The Internet embodies a new paradigm of distributed open computer networks, in which agents users and computational devices are not cooperative, but self-interested, with private
More informationModelling and Solving of Multidimensional Auctions 1
220 Ekonomický časopis, 65, 2017, č. 3, s. 220 236 Modelling and Solving of Multidimensional Auctions 1 Petr FIALA* 1 Abstract Auctions are important market mechanisms for the allocation of goods and services.
More informationcomputes the allocation and agent payments. Sealed-bid auctions are undesirable computationally because of this complete revelation requirement, which
Chapter 10 Conclusions Auctions oer great promise as mechanisms for optimal resource allocation in complex distributed systems with self-interested agents. However, limited and costly computation necessitates
More informationAN ANALYSIS OF LINEAR PRICES IN ITERATIVE COMBINATORIAL AUCTIONS
AN ANALYSIS OF LINEAR PRICES IN ITERATIVE COMBINATORIAL AUCTIONS Pavlo Shabalin, Alexander Pikovsky, Martin Bichler Internet-based Information Systems Roland Berger & O 2 Germany Chair, TU Munich {shabalin,
More informationRecap Beyond IPV Multiunit auctions Combinatorial Auctions Bidding Languages. Multi-Good Auctions. CPSC 532A Lecture 23.
Multi-Good Auctions CPSC 532A Lecture 23 November 30, 2006 Multi-Good Auctions CPSC 532A Lecture 23, Slide 1 Lecture Overview 1 Recap 2 Beyond IPV 3 Multiunit auctions 4 Combinatorial Auctions 5 Bidding
More informationOptimal Shill Bidding in the VCG Mechanism
Itai Sher University of Minnesota June 5, 2009 An Auction for Many Goods Finite collection N of goods. Finite collection I of bidders. A package Z is a subset of N. v i (Z) = bidder i s value for Z. Free
More informationChapter 10: Preference Elicitation in Combinatorial Auctions
Chapter 10: Preference Elicitation in Combinatorial Auctions Tuomas Sandholm and Craig Boutilier 1 Motivation and introduction The key feature that makes combinatorial auctions (CAs) most appealing is
More informationEnvy Quotes and the Iterated Core-Selecting Combinatorial Auction
Envy Quotes and the Iterated Core-Selecting Combinatorial Auction Abraham Othman and Tuomas Sandholm Computer Science Department Carnegie Mellon University {aothman,sandholm}@cs.cmu.edu Abstract Using
More informationCS364B: Frontiers in Mechanism Design Lecture #17: Part I: Demand Reduction in Multi-Unit Auctions Revisited
CS364B: Frontiers in Mechanism Design Lecture #17: Part I: Demand Reduction in Multi-Unit Auctions Revisited Tim Roughgarden March 5, 014 1 Recall: Multi-Unit Auctions The last several lectures focused
More informationcompute Vickrey payments. The basic auction procedure is described for an exclusive-or bidding language over items. However, as described in Section 5
Chapter 5 ibundle: An Iterative Combinatorial Auction In this chapter Iintroduce the ibundle ascending-price combinatorial auction, which follows quite directly from the primal-dual algorithm CombAuction
More informationChapter 1 Introduction
Chapter 1 Introduction There are many important settings in which multiple parties (or agents) must jointly make a decision to choose one outcome from a space of many possible outcomes based on the individuals
More informationOnline Ad Auctions. By Hal R. Varian. Draft: December 25, I describe how search engines sell ad space using an auction.
Online Ad Auctions By Hal R. Varian Draft: December 25, 2008 I describe how search engines sell ad space using an auction. I analyze advertiser behavior in this context using elementary price theory and
More informationPrices and Compact Bidding Languages for Combinatorial Auctions
Prices and Compact Bidding Languages for Combinatorial Auctions Robert W. Day Operations and Information Management, School of Business, University of Connecticut, Storrs, CT 06269-1041, Bob.Day@business.uconn.edu
More informationModeling of competition in revenue management Petr Fiala 1
Modeling of competition in revenue management Petr Fiala 1 Abstract. Revenue management (RM) is the art and science of predicting consumer behavior and optimizing price and product availability to maximize
More informationCS364B: Frontiers in Mechanism Design Lecture #11: Undominated Implementations and the Shrinking Auction
CS364B: Frontiers in Mechanism Design Lecture #11: Undominated Implementations and the Shrinking Auction Tim Roughgarden February 12, 2014 1 Introduction to Part III Recall the three properties we ve been
More informationTheoretical and Experimental Insights into Decentralized Combinatorial Auctions
Association for Information Systems AIS Electronic Library (AISeL) Wirtschaftsinformatik Proceedings 2011 Wirtschaftsinformatik 2011 Theoretical and Experimental Insights into Decentralized Combinatorial
More informationComputer Simulation for Comparison of Proposed Mechanisms for FCC Incentive Auction May 12, 2013
2013 Computer Simulation for Comparison of Proposed Mechanisms for FCC Incentive Auction May 12, 2013 By: Justin Huang Advisor: Professor Lawrence Ausubel University of Maryland, College Park Table of
More informationA Cooperative Approach to Collusion in Auctions
A Cooperative Approach to Collusion in Auctions YORAM BACHRACH, Microsoft Research and MORTEZA ZADIMOGHADDAM, MIT and PETER KEY, Microsoft Research The elegant Vickrey Clarke Groves (VCG) mechanism is
More informationAn Iterative Generalized Vickrey Auction: Strategy-Proofness without Complete Revelation
From: AAAI Technical Report SS-01-03. Compilation copyright 2001, AAAI (www.aaai.org). All rights reserved. An Iterative Generalized Vickrey Auction: Strategy-Proofness without Complete Revelation David
More informationMultiagent Systems: Spring 2006
Multiagent Systems: Spring 2006 Ulle Endriss Institute for Logic, Language and Computation University of Amsterdam Ulle Endriss (ulle@illc.uva.nl) 1 Combinatorial Auctions In a combinatorial auction, the
More informationVCG in Theory and Practice
1 2 3 4 VCG in Theory and Practice Hal R. Varian Christopher Harris Google, Inc. May 2013 Revised: December 26, 2013 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 It is now common to sell online ads using
More informationON QUADRATIC CORE PROJECTION PAYMENT RULES FOR COMBINATORIAL AUCTIONS YU SU THESIS
ON QUADRATIC CORE PROJECTION PAYMENT RULES FOR COMBINATORIAL AUCTIONS BY YU SU THESIS Submitted in partial fulfillment of the requirements for the degree of Bachelor of Science in Electrical and Computer
More informationCombinatorial auctions for electronic business
Sādhan ā Vol.30,Part2&3, April/June 2005, pp. 179 211. Printed in India Combinatorial auctions for electronic business Y NARAHARI 1 and PANKAJ DAYAMA 2 1 Electronic Enterprises Lab, Computer Science and
More informationThree New Connections Between Complexity Theory and Algorithmic Game Theory. Tim Roughgarden (Stanford)
Three New Connections Between Complexity Theory and Algorithmic Game Theory Tim Roughgarden (Stanford) Three New Connections Between Complexity Theory and Algorithmic Game Theory (case studies in applied
More informationMock Auction of LaGuardia Arrival/Departure Slots
Mock Auction of LaGuardia Arrival/Departure Lawrence M. Ausubel and Peter Cramton University of Maryland February 2005 Why Auctions? Efficiency Short run: resources to best use Long run: correct incentives
More informationFair Payments for Efficient Allocations in Public Sector Combinatorial Auctions
Fair Payments for Efficient Allocations in Public Sector Combinatorial Auctions Robert W. Day Operations and Information Management School of Business University of Connecticut Storrs, CT 06269-1041 S.
More informationCIS 700 Differential Privacy in Game Theory and Mechanism Design April 4, Lecture 10
CIS 700 Differential Privacy in Game Theory and Mechanism Design April 4, 2014 Lecture 10 Lecturer: Aaron Roth Scribe: Aaron Roth Asymptotic Dominant Strategy Truthfulness in Ascending Price Auctions 1
More informationVALUE OF SHARING DATA
VALUE OF SHARING DATA PATRICK HUMMEL* FEBRUARY 12, 2018 Abstract. This paper analyzes whether advertisers would be better off using data that would enable them to target users more accurately if the only
More informationCSC304: Algorithmic Game Theory and Mechanism Design Fall 2016
CSC304: Algorithmic Game Theory and Mechanism Design Fall 2016 Allan Borodin (instructor) Tyrone Strangway and Young Wu (TAs) November 2, 2016 1 / 14 Lecture 15 Announcements Office hours: Tuesdays 3:30-4:30
More informationdistributed optimization problems, such as supply-chain procurement or bandwidth allocation. Here is an outline of this chapter. Section 3.1 considers
Chapter 3 Computational Mechanism Design The classic mechanism design literature largely ignores computational considerations. It is common to assume that agents can reveal their complete preferences over
More informationPrototyping an Iterative Combinatorial Exchange
Prototyping an Iterative Combinatorial Exchange David C. Parkes Harvard University Airline Industry Wye River Workshop, June 23, 2004 Why Markets? Takeoff and landing slots are a constrained resource Markets
More informationDEFERRED ACCEPTANCE AUCTIONS AND RADIO SPECTRUM REALLOCATION
DEFERRED ACCEPTANCE AUCTIONS AND RADIO SPECTRUM REALLOCATION PAUL MILGROM AND ILYA SEGAL STANFORD AND AUCTIONOMICS 1 June 26, 2014 INCENTIVE AUCTIONS R&O September 28, 2012 F.C.C. BACKS PROPOSAL TO REALIGN
More informationThe Need for Information
The Need for Information 1 / 49 The Fundamentals Benevolent government trying to implement Pareto efficient policies Population members have private information Personal preferences Effort choices Costs
More informationThe Need for Information
The Need for Information 1 / 49 The Fundamentals Benevolent government trying to implement Pareto efficient policies Population members have private information Personal preferences Effort choices Costs
More information38 V. Feltkamp and R. Müller limitations of computability when their rules require the solution to NP-hard optimization problems. The complexity of su
The Infonomics Workshop on Electronic Market Design Vincent Feltkamp and Rudolf Müller International Institute of Infonomics The Infonomics Workshop on Electronic Market Design took place in Maastricht,
More informationValuation Uncertainty and Imperfect Introspection in Second-Price Auctions
Valuation Uncertainty and Imperfect Introspection in Second-Price Auctions Abstract In auction theory, agents are typically presumed to have perfect knowledge of their valuations. In practice, though,
More informationHow Does Computer Science Inform Modern Auction Design?
How Does Computer Science Inform Modern Auction Design? Case Study: The 2016-2017 U.S. FCC Incentive Auction Tim Roughgarden (Stanford) 1 F.C.C. Backs Proposal to Realign Airwaves September 28, 2012 By
More informationRobust Multi-unit Auction Protocol against False-name Bids
17th International Joint Conference on Artificial Intelligence (IJCAI-2001) Robust Multi-unit Auction Protocol against False-name Bids Makoto Yokoo, Yuko Sakurai, and Shigeo Matsubara NTT Communication
More informationOn Optimal Multidimensional Mechanism Design
On Optimal Multidimensional Mechanism Design YANG CAI, CONSTANTINOS DASKALAKIS and S. MATTHEW WEINBERG Massachusetts Institute of Technology We solve the optimal multi-dimensional mechanism design problem
More informationA Review of the GHz Auction
A Review of the 10-40 GHz Auction Peter Cramton 1 University of Maryland and Market Design Inc. For publication by Ofcom, September 2008 In February 2008, Ofcom s 10-40 GHz auction concluded. This was
More informationSearching for the Possibility Impossibility Border of Truthful Mechanism Design
Searching for the Possibility Impossibility Border of Truthful Mechanism Design RON LAVI Faculty of Industrial Engineering and Management, The Technion, Israel One of the first results to merge auction
More information1 Mechanism Design (incentive-aware algorithms, inverse game theory)
15-451/651: Design & Analysis of Algorithms April 10, 2018 Lecture #21 last changed: April 8, 2018 1 Mechanism Design (incentive-aware algorithms, inverse game theory) How to give away a printer The Vickrey
More informationAn Expressive Auction Design for Online Display Advertising
An Expressive Auction Design for Online Display Advertising The Harvard community has made this article openly available. Please share how this access benefits you. Your story matters Citation Lahaie,
More informationFederal Communications Commission ComCom Federal Office of Communications Worked Example - CCA
Federal Communications Commission ComCom Federal Office of Communications Worked Example - CCA Addendum to the invitation to tender for frequency blocks for the national provision of mobile telecommunication
More informationIncentive Auctions. Peter Cramton* Professor of Economics, University of Maryland Chairman, Market Design Inc. 23 May 2011 (updated 29 May 2011)
Incentive Auctions Peter Cramton* Professor of Economics, University of Maryland Chairman, Market Design Inc. 23 May 2011 (updated 29 May 2011) * Special thanks to Larry Ausubel, Evan Kwerel, and Paul
More informationOnline Combinatorial Auctions
Online Combinatorial Auctions Jan Ulrich Department of Computer Science University of British Columbia ulrichj@cs.ubc.ca December 27, 2006 Abstract This paper explores the feasibility and motivation for
More informationFirst-Price Auctions with General Information Structures: A Short Introduction
First-Price Auctions with General Information Structures: A Short Introduction DIRK BERGEMANN Yale University and BENJAMIN BROOKS University of Chicago and STEPHEN MORRIS Princeton University We explore
More information1 Mechanism Design (incentive-aware algorithms, inverse game theory)
15-451/651: Design & Analysis of Algorithms April 6, 2017 Lecture #20 last changed: April 5, 2017 1 Mechanism Design (incentive-aware algorithms, inverse game theory) How to give away a printer The Vickrey
More informationIncentive-Compatible, Budget-Balanced, yet Highly Efficient Auctions for Supply Chain Formation
Incentive-Compatible, Budget-Balanced, yet Highly Efficient Auctions for Supply Chain Formation Moshe Babaioff School of Computer Science and Engineering, The Hebrew University of Jerusalem, Jerusalem
More informationUniversity of California, Davis Date: June 24, PRELIMINARY EXAMINATION FOR THE Ph.D. DEGREE. Answer four questions (out of five)
University of California, Davis Date: June 24, 203 Department of Economics Time: 5 hours Microeconomics Reading Time: 20 minutes PREIMINARY EXAMINATION FOR THE Ph.D. DEGREE Answer four questions (out of
More informationMechanism Design in Social Networks
Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence (AAAI-17) Mechanism Design in Social Networks Bin Li, a Dong Hao, a Dengji Zhao, b Tao Zhou a a Big Data Research Center, University
More information1 Mechanism Design (incentive-aware algorithms, inverse game theory)
TTIC 31010 / CMSC 37000 - Algorithms March 12, 2019 Lecture #17 last changed: March 10, 2019 1 Mechanism Design (incentive-aware algorithms, inverse game theory) How to give away a printer The Vickrey
More informationIntro to Algorithmic Economics, Fall 2013 Lecture 1
Intro to Algorithmic Economics, Fall 2013 Lecture 1 Katrina Ligett Caltech September 30 How should we sell my old cell phone? What goals might we have? Katrina Ligett, Caltech Lecture 1 2 How should we
More informationPreference Elicitation and Query Learning
Preference Elicitation and Query Learning Avrim Blum 1, Jeffrey Jackson 2, Tuomas Sandholm 1, and Martin Zinkevich 1 1 Carnegie Mellon University, Pittsburgh PA 15213, USA, maz@cs.cmu.edu, WWW home page:
More informationApproximation and Mechanism Design
Approximation and Mechanism Design Jason D. Hartline Northwestern University May 15, 2010 Game theory has a great advantage in explicitly analyzing the consequences of trading rules that presumably are
More informationMultiagent Resource Allocation 1
Multiagent Resource Allocation 1 Michal Jakob Agent Technology Center, Dept. of Computer Science and Engineering, FEE, Czech Technical University AE4M36MAS Autumn 2014 - Lecture 11 Where are We? Agent
More informationReport 2: Simultaneous Ascending Auctions with Package Bidding
The opinions and conclusions expressed in this paper are those of the authors and do not necessarily reflect the views of the FCC or any of its Commissioners or other staff. CRA No. 1351-00 Report 2: Simultaneous
More informationChapter 17. Auction-based spectrum markets in cognitive radio networks
Chapter 17 Auction-based spectrum markets in cognitive radio networks 1 Outline Rethinking Spectrum Auctions On-demand Spectrum Auctions Economic-Robust Spectrum Auctions Double Spectrum Auctions for Multi-party
More informationCharacterization of Strategy/False-name Proof Combinatorial Auction Protocols: Price-oriented, Rationing-free Protocol
Characterization of Strategy/False-name Proof Combinatorial Auction Protocols: Price-oriented, Rationing-free Protocol Makoto Yokoo NTT Communication Science Laboratories 2-4 Hikaridai, Seika-cho Soraku-gun,
More informationThe Relevance of a Choice of Auction Format in a Competitive Environment
Review of Economic Studies (2006) 73, 961 981 0034-6527/06/00370961$02.00 The Relevance of a Choice of Auction Format in a Competitive Environment MATTHEW O. JACKSON California Institute of Technology
More informationComputationally Feasible VCG Mechanisms. by Alpha Chau (feat. MC Bryce Wiedenbeck)
Computationally Feasible VCG Mechanisms by Alpha Chau (feat. MC Bryce Wiedenbeck) Recall: Mechanism Design Definition: Set up the rules of the game s.t. the outcome that you want happens. Often, it is
More informationMotivated by the increasing use of auctions by government agencies, we consider the problem of fairly
MANAGEMENT SCIENCE Vol. 53, No. 9, September 2007, pp. 1389 1406 issn 0025-1909 eissn 1526-5501 07 5309 1389 informs doi 10.1287/mnsc.1060.0662 2007 INFORMS Fair Payments for Efficient Allocations in Public
More informationA Dynamic Programming Model for Determining Bidding Strategies in Sequential Auctions: Quasi-linear Utility and Budget Constraints
A Dynamic Programming Model for Determining Bidding Strategies in Sequential Auctions: Quasi-linear Utility and Budget Constraints Hiromitsu Hattori Makoto Yokoo Nagoya Institute of Technology Nagoya,
More informationAn Algorithm for Multi-Unit Combinatorial Auctions
From: AAAI-00 Proceedings Copyright 2000, AAAI (wwwaaaiorg) All rights reserved An Algorithm for Multi-Unit Combinatorial Auctions Kevin Leyton-Brown and Yoav Shoham and Moshe Tennenholtz Computer Science
More informationAuctions. José M Vidal. Department of Computer Science and Engineering University of South Carolina. March 17, Abstract
José M Vidal Department of Computer Science and Engineering University of South Carolina March 17, 2010 Abstract We introduce auctions for multiagent systems. Chapter 7. Valuations 1 Valuations 2 Simple
More informationQuadratic Core-Selecting Payment Rules for Combinatorial Auctions
OPERATIONS RESEARCH Vol. 60, No. 3, May June 2012, pp. 588 603 ISSN 0030-364X (print) ISSN 1526-5463 (online) http://dx.doi.org/10.1287/opre.1110.1024 2012 INFORMS Quadratic Core-Selecting Payment Rules
More informationGroup Incentive Compatibility for Matching with Contracts
Group Incentive Compatibility for Matching with Contracts John William Hatfield and Fuhito Kojima July 20, 2007 Abstract Hatfield and Milgrom (2005) present a unified model of matching with contracts,
More informationMarket Design and the Evolution of the Combinatorial Clock Auction
Market Design and the Evolution of the Combinatorial Clock Auction BY Lawrence M. Ausubel and Oleg V. Baranov* * Ausubel: Department of Economics, University of Maryland, 3114 Tydings Hall, College Park,
More informationSynergistic Valuations and Efficiency in Spectrum Auctions
Synergistic Valuations and Efficiency in Spectrum Auctions Andor Goetzendorff, Martin Bichler a,, Jacob K. Goeree b a Department of Informatics, Technical University of Munich, Germany goetzend@in.tum.de,
More informationCombinatorial Auctions
T-79.7003 Research Course in Theoretical Computer Science Phase Transitions in Optimisation Problems October 16th, 2007 Combinatorial Auctions Olli Ahonen 1 Introduction Auctions are a central part of
More informationEfficiency Guarantees in Market Design
Efficiency Guarantees in Market Design NICOLE IMMORLICA, MICROSOFT JOINT WORK WITH B. LUCIER AND G. WEYL The greatest risk to man is not that he aims too high and misses, but that he aims too low and hits.
More informationOptimizing Prices in Descending Clock Auctions
Optimizing Prices in Descending Clock Auctions Abstract A descending (multi-item) clock auction (DCA) is a mechanism for buying from multiple parties. Bidder-specific prices decrease during the auction,
More informationDo Core-Selecting Combinatorial Clock Auctions always lead to high Efficiency? An Experimental Analysis of Spectrum Auction Designs
Experimental Economics manuscript No. (will be inserted by the editor) Do Core-Selecting Combinatorial Clock Auctions always lead to high Efficiency? An Experimental Analysis of Spectrum Auction Designs
More informationMulti-Object Auctions with Package Bidding: An Experimental Comparison of ibea and Vickrey
Multi-Object Auctions with Package Bidding: An Experimental Comparison of ibea and Vickrey Yan Chen Kan Takeuchi October 10, 2005 Abstract We study two package auction mechanisms in the laboratory, a sealed
More informationUsing Expressiveness to Improve the Economic Efficiency of Social Mechanisms
Using Expressiveness to Improve the Economic Efficiency of Social Mechanisms Michael Benisch School of Computer Science, Carnegie Mellon University Joint work with: Norman Sadeh, Tuomas Sandholm 2 Talk
More information